SkBlurImageFilter.cpp revision 27eec46d6925afc76c2241c6b3182ce9b3284c9e 
1/*
2 * Copyright 2011 The Android Open Source Project
3 *
4 * Use of this source code is governed by a BSD-style license that can be
5 * found in the LICENSE file.
6 */
7
8#include "SkBitmap.h"
9#include "SkBlurImageFilter.h"
10#include "SkColorPriv.h"
11#include "SkFlattenableBuffers.h"
12#include "SkGpuBlurUtils.h"
13#include "SkBlurImage_opts.h"
14#if SK_SUPPORT_GPU
15#include "GrContext.h"
16#include "SkImageFilterUtils.h"
17#endif
18
19SkBlurImageFilter::SkBlurImageFilter(SkFlattenableReadBuffer& buffer)
20  : INHERITED(buffer) {
21    fSigma.fWidth = buffer.readScalar();
22    fSigma.fHeight = buffer.readScalar();
23    buffer.validate(SkScalarIsFinite(fSigma.fWidth) &&
24                    SkScalarIsFinite(fSigma.fHeight) &&
25                    (fSigma.fWidth >= 0) &&
26                    (fSigma.fHeight >= 0));
27}
28
29SkBlurImageFilter::SkBlurImageFilter(SkScalar sigmaX,
30                                     SkScalar sigmaY,
31                                     SkImageFilter* input,
32                                     const CropRect* cropRect)
33    : INHERITED(input, cropRect), fSigma(SkSize::Make(sigmaX, sigmaY)) {
34    SkASSERT(sigmaX >= 0 && sigmaY >= 0);
35}
36
37void SkBlurImageFilter::flatten(SkFlattenableWriteBuffer& buffer) const {
38    this->INHERITED::flatten(buffer);
39    buffer.writeScalar(fSigma.fWidth);
40    buffer.writeScalar(fSigma.fHeight);
41}
42
43enum BlurDirection {
44    kX, kY
45};
46
47/**
48 *
49 * In order to make memory accesses cache-friendly, we reorder the passes to
50 * use contiguous memory reads wherever possible.
51 *
52 * For example, the 6 passes of the X-and-Y blur case are rewritten as
53 * follows. Instead of 3 passes in X and 3 passes in Y, we perform
54 * 2 passes in X, 1 pass in X transposed to Y on write, 2 passes in X,
55 * then 1 pass in X transposed to Y on write.
56 *
57 * +----+       +----+       +----+        +---+       +---+       +---+        +----+
58 * + AB + ----> | AB | ----> | AB | -----> | A | ----> | A | ----> | A | -----> | AB |
59 * +----+ blurX +----+ blurX +----+ blurXY | B | blurX | B | blurX | B | blurXY +----+
60 *                                         +---+       +---+       +---+
61 *
62 * In this way, two of the y-blurs become x-blurs applied to transposed
63 * images, and all memory reads are contiguous.
64 */
65
66template<BlurDirection srcDirection, BlurDirection dstDirection>
67static void boxBlur(const SkPMColor* src, int srcStride, SkPMColor* dst, int kernelSize,
68                    int leftOffset, int rightOffset, int width, int height)
69{
70    int rightBorder = SkMin32(rightOffset + 1, width);
71    int srcStrideX = srcDirection == kX ? 1 : srcStride;
72    int dstStrideX = dstDirection == kX ? 1 : height;
73    int srcStrideY = srcDirection == kX ? srcStride : 1;
74    int dstStrideY = dstDirection == kX ? width : 1;
75#ifndef SK_DISABLE_BLUR_DIVISION_OPTIMIZATION
76    uint32_t scale = (1 << 24) / kernelSize;
77    uint32_t half = 1 << 23;
78#endif
79    for (int y = 0; y < height; ++y) {
80        int sumA = 0, sumR = 0, sumG = 0, sumB = 0;
81        const SkPMColor* p = src;
82        for (int i = 0; i < rightBorder; ++i) {
83            sumA += SkGetPackedA32(*p);
84            sumR += SkGetPackedR32(*p);
85            sumG += SkGetPackedG32(*p);
86            sumB += SkGetPackedB32(*p);
87            p += srcStrideX;
88        }
89
90        const SkPMColor* sptr = src;
91        SkColor* dptr = dst;
92        for (int x = 0; x < width; ++x) {
93#ifndef SK_DISABLE_BLUR_DIVISION_OPTIMIZATION
94            *dptr = SkPackARGB32((sumA * scale + half) >> 24,
95                                 (sumR * scale + half) >> 24,
96                                 (sumG * scale + half) >> 24,
97                                 (sumB * scale + half) >> 24);
98#else
99            *dptr = SkPackARGB32(sumA / kernelSize,
100                                 sumR / kernelSize,
101                                 sumG / kernelSize,
102                                 sumB / kernelSize);
103#endif
104            if (x >= leftOffset) {
105                SkColor l = *(sptr - leftOffset * srcStrideX);
106                sumA -= SkGetPackedA32(l);
107                sumR -= SkGetPackedR32(l);
108                sumG -= SkGetPackedG32(l);
109                sumB -= SkGetPackedB32(l);
110            }
111            if (x + rightOffset + 1 < width) {
112                SkColor r = *(sptr + (rightOffset + 1) * srcStrideX);
113                sumA += SkGetPackedA32(r);
114                sumR += SkGetPackedR32(r);
115                sumG += SkGetPackedG32(r);
116                sumB += SkGetPackedB32(r);
117            }
118            sptr += srcStrideX;
119            if (srcDirection == kY) {
120                SK_PREFETCH(sptr + (rightOffset + 1) * srcStrideX);
121            }
122            dptr += dstStrideX;
123        }
124        src += srcStrideY;
125        dst += dstStrideY;
126    }
127}
128
129static void getBox3Params(SkScalar s, int *kernelSize, int* kernelSize3, int *lowOffset,
130                          int *highOffset)
131{
132    float pi = SkScalarToFloat(SK_ScalarPI);
133    int d = static_cast<int>(floorf(SkScalarToFloat(s) * 3.0f * sqrtf(2.0f * pi) / 4.0f + 0.5f));
134    *kernelSize = d;
135    if (d % 2 == 1) {
136        *lowOffset = *highOffset = (d - 1) / 2;
137        *kernelSize3 = d;
138    } else {
139        *highOffset = d / 2;
140        *lowOffset = *highOffset - 1;
141        *kernelSize3 = d + 1;
142    }
143}
144
145bool SkBlurImageFilter::onFilterImage(Proxy* proxy,
146                                      const SkBitmap& source, const SkMatrix& ctm,
147                                      SkBitmap* dst, SkIPoint* offset) {
148    SkBitmap src = source;
149    if (getInput(0) && !getInput(0)->filterImage(proxy, source, ctm, &src, offset)) {
150        return false;
151    }
152
153    if (src.config() != SkBitmap::kARGB_8888_Config) {
154        return false;
155    }
156
157    SkAutoLockPixels alp(src);
158    if (!src.getPixels()) {
159        return false;
160    }
161
162    SkIRect srcBounds, dstBounds;
163    src.getBounds(&srcBounds);
164    if (!this->applyCropRect(&srcBounds, ctm)) {
165        return false;
166    }
167
168    dst->setConfig(src.config(), srcBounds.width(), srcBounds.height());
169    dst->getBounds(&dstBounds);
170    dst->allocPixels();
171    int kernelSizeX, kernelSizeX3, lowOffsetX, highOffsetX;
172    int kernelSizeY, kernelSizeY3, lowOffsetY, highOffsetY;
173    getBox3Params(fSigma.width(), &kernelSizeX, &kernelSizeX3, &lowOffsetX, &highOffsetX);
174    getBox3Params(fSigma.height(), &kernelSizeY, &kernelSizeY3, &lowOffsetY, &highOffsetY);
175
176    if (kernelSizeX < 0 || kernelSizeY < 0) {
177        return false;
178    }
179
180    if (kernelSizeX == 0 && kernelSizeY == 0) {
181        src.copyTo(dst, dst->config());
182        return true;
183    }
184
185    SkBitmap temp;
186    temp.setConfig(dst->config(), dst->width(), dst->height());
187    if (!temp.allocPixels()) {
188        return false;
189    }
190
191    const SkPMColor* s = src.getAddr32(srcBounds.left(), srcBounds.top());
192    SkPMColor* t = temp.getAddr32(0, 0);
193    SkPMColor* d = dst->getAddr32(0, 0);
194    int w = dstBounds.width(), h = dstBounds.height();
195    int sw = src.rowBytesAsPixels();
196    SkBoxBlurProc boxBlurX, boxBlurY, boxBlurXY;
197    if (!SkBoxBlurGetPlatformProcs(&boxBlurX, &boxBlurY, &boxBlurXY)) {
198        boxBlurX = boxBlur<kX, kX>;
199        boxBlurY = boxBlur<kY, kY>;
200        boxBlurXY = boxBlur<kX, kY>;
201    }
202
203    if (kernelSizeX > 0 && kernelSizeY > 0) {
204#ifndef SK_DISABLE_BLUR_DIVISION_OPTIMIZATION
205        boxBlurX(s,  sw, t, kernelSizeX,  lowOffsetX,  highOffsetX, w, h);
206        boxBlurX(t,  w,  d, kernelSizeX,  highOffsetX, lowOffsetX,  w, h);
207        boxBlurXY(d, w,  t, kernelSizeX3, highOffsetX, highOffsetX, w, h);
208        boxBlurX(t,  h,  d, kernelSizeY,  lowOffsetY,  highOffsetY, h, w);
209        boxBlurX(d,  h,  t, kernelSizeY,  highOffsetY, lowOffsetY,  h, w);
210        boxBlurXY(t, h,  d, kernelSizeY3, highOffsetY, highOffsetY, h, w);
211#else
212        boxBlurX(s,  sw, t, kernelSizeX,  lowOffsetX,  highOffsetX, w, h);
213        boxBlurY(t,  w,  d, kernelSizeY,  lowOffsetY,  highOffsetY, h, w);
214        boxBlurX(d,  w,  t, kernelSizeX,  highOffsetX, lowOffsetX,  w, h);
215        boxBlurY(t,  w,  d, kernelSizeY,  highOffsetY, lowOffsetY,  h, w);
216        boxBlurX(d,  w,  t, kernelSizeX3, highOffsetX, highOffsetX, w, h);
217        boxBlurY(t,  w,  d, kernelSizeY3, highOffsetY, highOffsetY, h, w);
218#endif
219    } else if (kernelSizeX > 0) {
220        boxBlurX(s,  sw, d, kernelSizeX,  lowOffsetX,  highOffsetX, w, h);
221        boxBlurX(d,  w,  t, kernelSizeX,  highOffsetX, lowOffsetX,  w, h);
222        boxBlurX(t,  w,  d, kernelSizeX3, highOffsetX, highOffsetX, w, h);
223    } else if (kernelSizeY > 0) {
224        boxBlurY(s,  sw, d, kernelSizeY,  lowOffsetY,  highOffsetY, h, w);
225        boxBlurY(d,  w,  t, kernelSizeY,  highOffsetY, lowOffsetY,  h, w);
226        boxBlurY(t,  w,  d, kernelSizeY3, highOffsetY, highOffsetY, h, w);
227    }
228    offset->fX += srcBounds.fLeft;
229    offset->fY += srcBounds.fTop;
230    return true;
231}
232
233bool SkBlurImageFilter::filterImageGPU(Proxy* proxy, const SkBitmap& src, const SkMatrix& ctm,
234                                       SkBitmap* result, SkIPoint* offset) {
235#if SK_SUPPORT_GPU
236    SkBitmap input;
237    if (!SkImageFilterUtils::GetInputResultGPU(getInput(0), proxy, src, ctm, &input, offset)) {
238        return false;
239    }
240    GrTexture* source = input.getTexture();
241    SkIRect rect;
242    src.getBounds(&rect);
243    if (!this->applyCropRect(&rect, ctm)) {
244        return false;
245    }
246    SkAutoTUnref<GrTexture> tex(SkGpuBlurUtils::GaussianBlur(source->getContext(),
247                                                             source,
248                                                             false,
249                                                             SkRect::Make(rect),
250                                                             true,
251                                                             fSigma.width(),
252                                                             fSigma.height()));
253    offset->fX += rect.fLeft;
254    offset->fY += rect.fTop;
255    return SkImageFilterUtils::WrapTexture(tex, rect.width(), rect.height(), result);
256#else
257    SkDEBUGFAIL("Should not call in GPU-less build");
258    return false;
259#endif
260}
261